IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH ......Lafayette, IN 47907 USA (e-mail: [email protected]). N. Pierdicca is with the Department of Information Engineering, Electronics

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING 1

InSAR Water Vapor Data Assimilation intoMesoscale Model MM5: Technique and Pilot Study

Emanuela Pichelli, Rossella Ferretti, Domenico Cimini, Giulia Panegrossi, Daniele Perissin,Nazzareno Pierdicca, Senior Member, IEEE, Fabio Rocca, and Bjorn Rommen

Abstract—In this study, a technique developed to retrieve inte-grated water vapor from interferometric synthetic aperture radar(InSAR) data is described, and a three-dimensional variationalassimilation experiment of the retrieved precipitable water vaporinto the mesoscale weather prediction model MM5 is carried out.The InSAR measurements were available in the framework of theEuropean Space Agency (ESA) project for the “Mitigation of elec-tromagnetic transmission errors induced by atmospheric watervapor effects” (METAWAVE), whose goal was to analyze and pos-sibly predict the phase delay induced by atmospheric water vaporon the spaceborne radar signal. The impact of the assimilation onthe model forecast is investigated in terms of temperature, watervapor, wind, and precipitation forecast. Changes in the modeleddynamics and an impact on the precipitation forecast are found.A positive effect on the forecast of the precipitation is found forstructures at the model grid scale or larger (1 km), whereas a neg-ative effect is found on convective cells at the subgrid scale thatdevelops within 1 h time intervals. The computation of statisticalindices shows that the InSAR assimilation improves the forecast ofweak to moderate precipitation (<15 mm/3 h).

Index Terms—Atmospheric path delay, data assimilation,numerical weather prediction (NWP), synthetic aperture radar(SAR), water vapor.

I. INTRODUCTION

O NE OF THE major error sources in the short-term fore-cast of precipitation is the lack of precise and continuous

measurements of water vapor data [1], [2]. The water vapor is

Manuscript received February 13, 2014; revised July 04, 2014; acceptedSeptember 01, 2014. This work was supported in part by the European SpaceAgency under contract N. ESTEC-21207/07/NL/HE.

E. Pichelli and R. Ferretti are with the Department of Physics andChemistry, CETEMPS, University of L’Aquila, 67100 L’Aquila, Italy (e-mail:[email protected]; [email protected]).

D. Cimini is with the Institute of Methodologies for Environment Analysis,CNR, 85050 Potenza, Italy, and also with the Department of Physics andChemistry, CETEMPS, University of L’Aquila, 67100 L’Aquila, Italy (e-mail:[email protected]).

G. Panegrossi is with the Institute of Atmospheric Sciences and Climate,CNR, 00133 Rome, Italy (e-mail: [email protected]).

D. Perissin is with the School of Civil Engineering, Purdue University, WestLafayette, IN 47907 USA (e-mail: [email protected]).

N. Pierdicca is with the Department of Information Engineering, Electronicsand Telecommunications, Sapienza University of Rome, 00184 Rome, Italy(e-mail: [email protected]).

F. Rocca is with the DEI, Politecnico di Milano, 20133 Milan, Italy (e-mail:[email protected]).

B. Rommen is with the European Space Research and Technology Centre,European Space Agency (ESTEC/ESA), 2200AG Noordwijk, The Netherlands(e-mail: [email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/JSTARS.2014.2357685

an extremely important element of the atmosphere because itsdistribution is related to clouds, precipitation formation, and itrepresents a large proportion of the energy budget in the atmo-sphere. Its representation inside numerical weather prediction(NWP) models is critical to improve the weather forecast. It isalso very challenging because water vapor is involved in pro-cesses over a wide range of spatial and temporal scales. Animprovement in atmospheric water vapor monitoring that canbe assimilated in NWP models would improve the forecastaccuracy of precipitation and severe weather [1], [3]. In thisframework, the spaceborne interferometric synthetic apertureradar (InSAR), a useful tool for high-resolution water vaporretrieval [4], represents an interesting source of data to beassimilated into mesoscale models. Panegrossi et al. [5] havedemonstrated the InSAR capability of providing soil moisturemaps to constrain the surface boundary conditions in NWPmodels. InSAR is based on the measurement of the phase dif-ferences, associated with the distance between the satellite andeach land surface element, as observed from different satel-lite positions or at different times [6]. The neutral atmosphereintroduces an unknown delay in the SAR signal propagation,particularly, due to the high water vapor spatial and temporalvariability. Due to the differential nature of the InSAR tech-nique, the tropospheric contribution to the SAR interferogram,i.e., the so-called atmospheric phase screen (APS), is actu-ally related to the difference of delays due to the atmospherewhen SAR signals propagate through it, rather than their abso-lute value [7]. This effect can be exploited, offering a potentialsource of integrated water vapor (IWV) data (i.e., precipitablewater) with a high spatial resolution, provided that the groundmotion and topography effects are removed to isolate the watervapor contribution [3].

This paper presents a numerical experiment carried out bya variational data assimilation system (3DVAR) to assimilateInSAR observations into the Pennsylvania State Universitymesoscale model MM5 and to test their impact on the fore-cast. The InSAR data used in the assimilation were collectedduring the 2008 campaign of the ESA project METAWAVE(Mitigation of electromagnetic transmission errors inducedby atmospheric water vapor effects) [8]. In this project, theeffect of water vapor path delay in InSAR applications wasdeeply investigated, trying to identify and compare possibleindependent sources of information valuable to mitigate thoseartifacts [3], [8]–[13]. Several methods and tools have beenexploited in order to retrieve the water vapor field and relatedcharacteristics at resolution suitable for mitigating its effecton InSAR.

1939-1404 © 2014 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistributionrequires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.


2 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

In this work, we try to turn the InSAR tropospheric noiseinto an opportunity for NWP models. A major difficulty is thedifferential nature of the APS data (in time and space). APS pro-vides high-resolution mapping of the atmospheric path delaychanges with time where the earth surface remains steady, butthey do not furnish absolute values. This is a relevant drawbackif APSs are sampled at long time intervals, as in the case of theradar aboard the Envisat satellite, thus preventing, for instance,four-dimensional (4-D) variational assimilation. However, therevisit time is going to increase to the order of few days (as forCOSMO-SkyMed and Sentinel 1) thanks to the emerging con-stellations of radar sensors and could even become to the orderof hours considering the possibility to deploy a SAR aboard ageostationary orbit [14].

The goal of this study is to verify if the assimilation ofIWV data retrieved from InSAR have an impact on numeri-cal weather forecasts, and to be able to assess if the impact ispositive and if the precipitation forecast improves.

Previous studies demonstrated that the assimilation of IWVleads to better initial conditions (ICs) and that the retrievalof vertical structure of water vapor from observed data canimprove the precipitation forecast, especially if it is associatedwith the assimilation of wind observations [2].

The lack of the absolute value of IWV data using APS canbe partially solved by merging the high-resolution differentialinformation with a smoothed background provided by a statisti-cal analysis of water vapor map temporal series. In this way, thehigh-resolution APS information, estimated from the AdvancedSAR (ASAR) aboard of the Envisat satellite (by using thePermanent Scatterer (PS) multipass technique [15]), can beused to NWP models’ ICs. The methodology used to face thisproblem and the obtained results are described in Section II.

Section III presents an overview of the case study (October 3,2008) that corresponds to one of the two overpasses of Envisatcollected during the 15-day METAWAVE campaign. Differentmeasurements were available for comparison and validation ofthe results, including vertical profiles of thermodynamical vari-ables and radar data of reflectivity and precipitation. Unstableconditions of that day allowed to evaluate the impact of theassimilation also on the precipitation, one of the most difficultfield to be predicted.

The NWP model configuration and a brief summary of theassimilation method are described in Section IV. Results ofthe assimilation experiment and their discussion are presentedin Section V. The model results after assimilation of IWVretrieved from APSs are compared to a control simulation(without any kind of assimilation) to evaluate its impact onthe simulation. Radio soundings, meteorological radar, and raingauges observations are used as reference. The conclusion ofthis study is summarized in Section VII.

II. IWV MAPS FROM INSAR

The first problem to be solved to assimilate InSAR APS datais to obtain the absolute IWV value. The atmospheric delaymeasured by the SAR has been scaled according to its obser-vation angle (19.2◦–26.7◦) to find correspondence with the

vertical profiles of MM5, and a zenith-equivalent atmosphericdelay is considered hereafter.

The derivation of absolute IWV from InSAR is not straight-forward, as an interferogram from SAR is proportional to thepath delay difference in both time and space [16], and theInSAR APS at any given time is obtained by removing theuncorrelated noise and a phase ramp (i.e., a bilinear function ofthe image coordinates which can originate from satellite orbiterrors). Beside the conversion from path delay to IWV, whichcan be assumed roughly based on a proportionality factor [13],the basic idea is to estimate the time average distribution ofIWV in a given area and within a time frame by relying on anexternal source, such as IWV maps from other earth observation(EO) sensors.

The interferometric phase at a point r = [x, y] in the imagerepresents the phase difference between two SAR overpasses(at time i and j, respectively) referred to a single point r0. Itcontains a surface displacement term (subscript DISPL) anda term due to atmospheric path delay distortions (subscriptATMO). The image acquisition’s time difference is indicatedby Δij , Φ denotes the phase, and L is the path delay; for a SARinterferogram, we can write

ΔijΦ(r)−ΔijΦ(r0) = ΔijΦDISPL(r)

+4π

λΔijLATMO(r)−ΔijΦ(r0)

(1)

where the one-way atmospheric path delay L and phase are sim-ply related by the wave number k = 2π/λ multiplied by 2 toconsider the radar signal round-trip, where λ is the wavelength.In clear sky conditions, the path delay L has a hydrostatic dryterm and a term which is approximately proportional to IWVthrough a factor Π (i.e., L = IWV/Π). It can be demonstratedthat this factor is roughly 0.15 (slightly depending on condi-tions), so that 1 mm of IWV corresponds to roughly 6 mm ofpath delay due to water vapor [13].

For a nondeforming earth surface or a surface whose defor-mation can be modeled, ΔijΦDISPL can be removed. Theatmospheric delay disturbance (or APS) can be produced ineach point r by differencing a sequence of SAR images withrespect to a master image taken at time j = M, with an arbitraryunknown constant (consti) corresponding to the phase differ-ence of the reference point x0, i.e., consti = λ ·ΔΦiM (r0)/4π

APSi(r) = Li(r)− LM (r)− consti. (2)

In (2), we can assume that the hydrostatic component is con-stant within a typical SAR image (i.e., it is assumed that thereare no significant changes in the surface pressure and temper-ature fields within the extension of the image which is lessthan 100 km). The hydrostatic component is thus included intoconsti, and the remaining path delay is mainly related to IWV.

However, the decreasing path delay over mountain sites, dueto the shorter distance traveled through the atmosphere, is alsorelated to the vertical stratification of the atmosphere [13].Basili et al. [13] show that the trend of wet path delay withrespect to surface height ranges from 1 cm/km for a dry atmo-sphere to more than 5 cm/km for a very moist atmosphere. The


PICHELLI et al.: InSAR WATER VAPOR DATA ASSIMILATION INTO MESOSCALE MODEL MM5 3

dry path delay also contributes to this trend, but its changes areassumed to be much lower.

Thus, in order to derive an absolute value of the path delay ata given time i from the APS, one should know the atmosphericconditions at the time of the master image acquisition (i.e.,LM ), though the ambiguity associated with consti remains.LM could be provided by other sources, e.g., the EO sen-sors sensitive to water vapor content, either in the infrared ormicrowave spectral band. The NWP outputs could provide LM

to estimate Li, but these cannot be used when attempting toassimilate the absolute APS, as it would introduce statisticallycorrelated observations into the assimilation process. Thus, theerror variance associated with the water vapor content of themaster acquisition provided by the external source (σ2

EXT) canbe significant. This would add up to the APS intrinsic error(σ2

APS) resulting in the error variance of the estimated IWV attime i (σ2

IWV )

σ2IWVi

= σ2EXT + σ2

APSi. (3)

If a suitable sequence of external data was available, anotherapproach could be followed by averaging many APS imagesand relying on the external source to estimate the expectedvalue of the atmosphere path delay (assuming that the SARand the external sources are observing on average the sameatmosphere). We would have

Mean [APSi(r)] = Mean[LEXTi (r)

]− LM (r)− const

from which we can estimate the master to be substituted into(2), and thus more reliably estimate the absolute atmosphericdelay from APS as

Li(r) = APSi(r)− Mean [APSi(r)] + Mean[LEXTi (r)

]

+ consti. (4)

Note that for each time i, there is still an unknown constantwhich can be estimated from the external source at that specifictime, as it is explained below. The advantage of (4) is in thelower influence of the external source errors, as the error vari-ance of the mean estimates is reduced by a factor equal to thenumber of available observations used in the averaging process,which should be very large

σ2IWVi

= σ2Mean EXT + σ2

Mean APS + σ2APSi

. (5)

In our experiment, the external source to estimate the IWVcomes from the MEdium Resolution Imaging Spectrometer(MERIS) aboard Envisat, which provides images simultane-ously to the ASAR acquisitions. Note that clouds affect IWVestimation from MERIS; therefore, the MERIS operationalcloud mask is used for discarding cloud-contaminated pixels. Inspite of this, when performing the averaging of many APS mapsrequired in (4), the mean background maps become availableeverywhere in the SAR frame (assuming clear-sky conditionsfor at least few cases in the stack). From the absolute path delayretrieved as in (4), the IWV is finally derived

IWVAPSi (r) = ΠLi(r) = Π(λ/4π)Φi(r). (6)

Note that (6) is still wrapped, hence the 2π phase ambigu-ity affecting the APS should be added. Since ASAR works at5.3 GHz, in this specific case, the 2π phase ambiguity cor-responds to a variation in the slant path delay that folds onewavelength λ = 5.66 cm; considering that the observing angleis θz = 19.2◦ − 26.7◦ and that the SAR signal travels the atmo-sphere twice, this slant path delay corresponds to a variationin IWV equal to dIWV = λ/2 sec(θz) ∗Π ≈ 0.40± 0.01 cm.Thus, the APS 2π phase ambiguity maps into a IWV ambigu-ity approximately equal to dIWV2π ∼ 0.40 cm. However, APSmay also be affected by a phase ramp, which consists of a phaseresidual associated with orbital errors showing a linear trendwith the position on a horizontal plane x, y (i.e., East and Northcoordinates in the map). If information about the IWV fieldis coming from an external source (e.g., MERIS in our case),the constant term, the 2π phase ambiguity, and the phase rampcan be removed. A simple two-step process is used to removethese terms: 1) search for the two-dimensional (2-D) east/northplanar trend in the difference between IWVEXT

i and IWVAPSi

(assuming consti = 0) and 2) remove the obtained plane fromIWVAPS

i , i.e.,

IWVAPSi = Π {APSi(r)− Mean [APSi(r)]}

+ Mean[IWVEXT

i (r)]+ F i(x, y) (7)

F i(x, y) = aio + aix · x+ aiy · y[aio, a

ix, a

iy

]= FIT2D

{IWVEXT

i (r)−Π · {APSi(r)

− Mean [APSi(r)]} − Mean[IWVEXT

i (r)]}

(8)

where FIT2D represents a two-dimentional least square fittingoperator.

However, an extensive cloudiness can reduce the numberof pixels with a meaningful IWV, and thus it could precludethe trend estimation. In addition, MERIS IWV may appearunder/overestimated in the presence of undetected cloudy pix-els (e.g., at cloud edges) or pixels under cloud shadows, respec-tively [17]. Some measures have been adopted to reduce the riskof affecting the trend estimation. The most stringent cloud maskprovided by MERIS has been considered for discarding all thepixels that are detected as cloudy or undetermined at any rate.This eliminates the pixels where the IWV is not provided, andit also reduces the chances for cloud edges and cloud shadow.Then, the trend estimation is applied only if MERIS data areplenty (> 50% of the domain area) and approximately evenlydistributed.

In conclusion, in this approach, the APS brings informationon the high spatial frequency component of the path delay,whereas the low frequency component still needs to be pro-vided by other sources of information (like independent modelanalysis or EO products, such as MERIS in our case).

III. OVERVIEW OF THE CASE STUDY

Within the framework of the METAWAVE project, a compar-ison between ENVISAT interferograms and NWP model runs



was performed for a period of several years for the area of Rome(Central Italy) and Como (Northern Italy).

During Autumn 2008, the experimental campaign took placeand lasted for about 15 days. Microwave radiometers, radiosoundings, and GPS receivers were deployed both in Rome andComo areas, and data from available satellites and GPS receiveroperational networks were regularly collected to perform acomparison exercise [11]. During that period, Envisat over-passed the experiment area in Rome twice, collecting images,respectively, from an ascending and descending orbit. One ofthe days of the experimental campaign is used for this study:October 3, 2008, for its unstable meteorological conditions,represents a significant test case for studying the assimilationimpact on precipitation.

During October 3, 2008, a cold front associated with a NorthAtlantic cyclone crossed over Italy, followed by an anticycloneentering from the west side of the Mediterranean basin (notshown). The radio soundings in Pratica di Mare (South westof Rome, 41.65◦N, 12.43◦E) showed a weakly unstable atmo-sphere at 00 UTC (Fig. 1, top panel) with south-westerly windsat the surface and westerly winds at upper levels. The instabil-ity increased in the following hours and a south-southwesterlywind component was detected, as shown in the radio soundingof 12 UTC of the same day (Fig. 1, bottom panel). The incom-ing cold air mass contributed to increase the humidity of themiddle atmosphere during the day (in Fig. 1, the dew point tem-perature at 12 UTC is closer to the temperature curve than in theprevious profile in the layer between 850 and 600 hPa) increas-ing its instability [the convective available potential energy(CAPE) grows from almost 16.7 J/kg up to 67.7 J/kg]. Thedecrease of the lifted index (LIFT) and the increase of theK-index (KINX) indicate the increasing probability of widelyscattered thunderstorms occurrence over the region.

The Doppler radar located on the Midia mountain (42.05◦N,13.17◦E) recorded echoes from 12 UTC off the coast south-west of Rome (not shown); in the following 2 h, scatteredcells with reflectivity between 15 and 35 dBz (equivalent torainfall rate up to 6 mm/h) were detected over most of thesouthern part of Lazio [between the cities of Latina (LT) andFrosinone (FR), top left panel of Fig. 2]; after 16 UTC pre-cipitation was observed also in the innermost territory east ofRome (Fig. 2, top right panel). Some localized cells exceeding35 dBz were found at 17 UTC (not shown). Finally, moder-ate to locally heavy rain cells were detected between 20 and21 UTC (Fig. 2, bottom panels), with localized maxima of40–45 dBz (12–24 mm/h of rainfall); after this time, raingradually decreased, ending by midnight.

IV. MODEL CONFIGURATION AND DATA

ASSIMILATION TECHNIQUE

A. MM5 Configuration

The fifth generation of National Center for AtmosphericResearch (NCAR) and Pennsylvania State University (PSU)mesoscale model (MM5) is used in this study. This is a nonhy-drostatic model at primitive equations with a terrain-followingvertical coordinate and multiple nesting capabilities [18]. Fourtwo-way nested domains are used to simulate the weather

Fig. 1. Radio soundings from Pratica di Mare, at the center of the coast ofLazio, Central Italy, at 00 UTC of October 3, 2008 (top) and at 12 UTC(bottom). The two black lines represent, respectively, the dew point tempera-ture (◦C, left) and the temperature (◦C, right). Wind barbs are plotted on theright (Data available on weather.uwyo.edu).

event on October 3, 2008 (Fig. 3) to be able to enhance thehorizontal resolution over the urban area of Rome. The outerdomain covers most of western Mediterranean area, centeredat 41.5◦N, 10.0◦E with 27 km spatial resolution (D01 inFig. 3). The nested domains cover Central Italy with a spatialresolution of 9 km for domain 2, 3 km for domain 3, and 1 kmfor the innermost domain (D04 in Fig. 3). D04 covers the cityarea and its surroundings (Lazio region) and it overlaps theERS satellite swath.

The following model configuration has been used, based onprevious studies and sensitivity test over the same area [19]:

1) 33 unequally spaced vertical sigma levels (σ), from thesurface up to 100 hPa, with a higher resolution in the plan-etary boundary layer (PBL) than in the free atmosphere;

2) the medium-range forecast MRF scheme for the PBL.This scheme is based on the Troen-Mahrt representationof counter-gradient term and the eddy viscosity profile inthe well-mixed PBL [20];



Fig. 2. Reflectivity maps on October 3, 2008 measured by the Dopplerradar located over Midia Mountain (Central Italy), owned by the NationalDepartment of Civil Protection of Italy.

Fig. 3. MM5 domains configuration. Domain D01 has resolution of 27 km;D02 has resolution of 9 km; D03 has resolution of 3 km; and D04 has resolutionof 1 km.

3) the CLOUD radiation scheme for radiative transferprocesses. This scheme accounts for both shortwave andlongwave interactions with explicit cloud and clear-airscattering [21];

4) the Kain-Fritsch-2 cumulus convection parameterizationis used for domains 1 and 2 [22], [23], whereas nocumulus scheme is used for domains 3 and 4;

5) the Reisner-2 scheme for microphysics; based on mixed-phase scheme, graupel, and ice number concentrationprediction equations [24].

The European Centre for Medium-Range Weather Forecast(ECMWF) analysis at 0.25◦ horizontal resolution for tempera-ture, wind speed, relative humidity, and geopotential height is

interpolated to the MM5 horizontal grid and to sigma levels toproduce the model initial and boundary conditions.

B. InSAR Data Assimilation

The atmospheric data assimilation aims to incorporate obser-vations into NWP models and to fill data gaps using physical,dynamical, and/or statistical information. Physical consistency,spatial and temporal coherence, and noise suppression are threeof the major concerns in atmospheric data assimilation.

Briefly, the variational method is an optimization problem:3DVAR attempts to find the best fit of a gridded representationof the state of the atmosphere (first guess or background field) toa discretely and irregularly distributed set of observations [25],[26]. The best fit is obtained by minimizing the so-called costfunction J, defined as

J = Jb + Jo =1

2

(xb − x

)TB−1

(xb − x

)

+1

2

(yo −H

(xb

))T(E + F )−1

(yo −H

(xb

))(9)

where xb is the background term, yo is the generic observation,H(xb) is the corresponding value evaluated by the operatorH used to transform the gridded analysis into the observationspace. The solution of this equation x = xa is the a posteriorimaximum likelihood estimate of the true state of the atmo-sphere; B, E, and F are the covariance error matrices for thebackground, the observations, and the operator H, respectively.

The 3DVAR is used to assimilate data of IWV, retrieved fromInSAR and an external data source (MERIS) (as described inSection II), with the aim of improving ICs for MM5 [27].

Four interferometric stacks of ASAR images, acquired by theC-band radar aboard the European Envisat satellite in standardStripmap mode, with a single-look resolution of 9 by 6 m (slantby azimuth), and over a swath of about 100 km, have been pro-cessed using the PS technique to generate InSAR APSs. Theone used in this study, formed by 10 images, was collected overRome along the Envisat descending track 351.

The InSAR APS measurements of the atmospheric pathdelay can be assumed similar to data provided by a dense GPSreceivers network, thus the H operator implemented for GPS[28], [29] has been adopted in (9).

The descending SAR overpass was acquired at 0930 UTC onOctober 3, 2008, therefore a background analysis (xb) is nec-essary at that time to assimilate the related APS data. StandardECMWF analysis usually used to initialize model simulationsis available every 6 h at synoptic times (e.g., 06 or 12 UTC),far away from the SAR overpass. A short-term MM5 simula-tion starting at 06 UTC of October 3 and ending at 09 UTCof the same day has been produced; the output at 09 UTC hasbeen fed back to MM5 to be used as first guess for the 3DVAR.This procedure allows for having ICs at 09 UTC to initialize theforecast with assimilated InSAR data (MM5_VAR). Similarly,IC without assimilation of InSAR data is produced to performa control run (MM5_NOVAR). No change of the atmosphereconditions is assumed between 09 UTC and 0930 UTC in orderto use the ENVISAT data acquired at 0930 UTC (hypothesis offrozen atmosphere).



Fig. 4. Integrated water vapor at simulation start time (00 UTC of October 3, 2008) by MM5_NOVAR (left panel) and by InSAR down-sampled at the model gridresolution (right panel).

Moreover, a background (B) and an observation error matrix(E) need to be defined, as shown in (9). The B matrix isrelated to the climatology of the event and it is calculated onthe whole month of October. To compute the B matrix, the“NMC method” is commonly used for NWP models [30], [31],where the forecast error covariance is calculated by using fore-cast difference statistics (e.g., differences between forecastsat T+ 48 and at T+ 24). The E matrix is built based uponthe assumptions of 1) a constant IWV error estimated withinσAPS = 0.05 cm [see (5)], which corresponds to a random erroron the InSAR phase of the order of 15◦, and 2) no cross corre-lation of observation errors between adjacent pixels. To makecondition 2) as true as possible, a thinning of the observa-tions is performed. Cardinali et al. [32] demonstrated that theinfluence of the assimilated data in the variational assimilationprocess is lower in data-rich areas and that a large error corre-lation among them decreases the observation influence in theassimilation process, increasing the weight of the backgroundfield [32]. Thus, the whole set of InSAR data has been down-sampled to the resolution of the innermost domain (1 km): foreach MM5 grid point, the nearest available InSAR measure-ment is retained. An alternative to the thinning process wouldbe an iterative cycle of assimilation which would allow to getonly a portion of data at each assimilation cycle [32], but thismethod will be considered for future work.

V. RESULTS OF THE DATA ASSIMILATION EXPERIMENT

A. Impact on ICs

The assimilation procedure of any meteorological fieldrequires the adjustment of the other input variables (tem-perature, pressure, winds, etc.) in order to be coherent withthe changes deriving from 3DVAR on the assimilated field.Fig. 4 shows the IWV field at MM5 start time when noassimilation is performed (MM5_NOVAR, left panel) and theIWV retrieved from InSAR data after thinning process and

successively assimilated into the model IC (right panel). TheInSAR data show a larger variability than MM5 in the same areaand moister conditions along the Tiber Valley around Rometoward the coastline.

The impact of the water vapor assimilation on the IC canbe evaluated by analyzing the increments with respect to thebackground field. An example is given in Fig. 5, showing thewater vapor mixing ratio (QVP, left panel) and the ground tem-perature (TGK, right panel) increments over MM5 innermostdomain at the start time (09 UTC).

The QVP increments (Fig. 5, left panel) clearly show thatchanges occur in the area where the InSAR data are available,whereas the temperature adjustments (Fig. 5, right panel) arespread all over the domain. The increments of QVP range from0% to around the 4% of the first guess field (MM5_NOVAR),whereas surface temperature increments rise up to a maximumof 6%.

An assimilation experiment with MERIS data within theEnvisat swath (MM5_ME) was also performed (not shown) toevaluate their impact on the final results, as MERIS has beenused for IWV retrieval from InSAR data (Section II). It wasfound that mainly negative increments are produced by MERISassimilation on QVP field; also in this case, increments arenonzero only where MERIS data are available. At the starttime, the IWV of the MM5_VAR simulation shows a largerspatial variability than MM5_ME and an IWV increase close tothe coastline that is not present on MM5_ME. A comparison ofMM5_ME with observations, analogous to the one that is pre-sented in the following sections, shows profiles similar to thoseproduced by InSAR assimilation but with larger biases formost of the variables. On the other hand, the comparison showsa negligible impact on the precipitation field. This impliesthat the enhanced spatial variability introduced by the InSARis crucial to produce changes that will be discussed below.A deeper investigation of these results would be necessary,but they are sufficient to ascribe the improvements found



Fig. 5. Increment of water vapor mixing ratio (g/kg) at 1000 hPa (left panel) pressure level and of the ground temperature (K, right panel) at start time 09 UTCon the highest resolution domain of MM5 model after InSAR data assimilation. AB (black) and CD (red) are two cross-sectional lines.

for MM5_SAR mainly to the InSAR data assimilation, asdiscussed later.

B. Vertical Structure: Water Vapor and Soundings

As a first assessment of the impact of InSAR assimila-tion, a comparison of the vertical distribution of water vaporat the lower atmospheric layers between the two simulations(MM5_NOVAR and MM5_VAR) has been performed. A ver-tical cross section centered in Rome and crossing the InSARswath (Fig. 5, line AB) on MM5 domain 4 shows that thelargest differences between MM5_NOVAR and MM5_VARsimulations are seen within 3 h from the start time, whereasthey are negligible after 12 UTC. Between 09 UTC and 10UTC, the two simulations show differences both in the ver-tical content of water vapor and in its horizontal variability(Figs. 6 and 7), whereas only small differences on its con-tent are found in the following hours. Fig. 6 shows the crosssection at 09 UTC of October 3 for the two simulations(MM5_NOVAR on the left and MM5_VAR on the right); theassimilation of InSAR data (right panel) causes both a reduc-tion of the vapor content and a cooling of the layers: the9.6 g/kg contour, e.g., reaches 750 m instead of 830 m as forMM5_NOVAR.

These characteristics are found all along the vertical crosssection. By 10 UTC, changes in the vertical section are foundespecially across the urban area of Rome (area inside the twogray dashed lines in Fig. 7). A few differences are found alsoabove 1 km for both the water vapor and the thermal structure,and will be discussed later.

This first comparison allows to assess an impact of the assim-ilation on the evolution of the atmospheric conditions, but afurther and more objective comparison with experimental datais performed to evaluate its effectiveness.

A comparison between model results and radio-soundingobservations (RAOB) is performed to investigate the InSARdata impact on the profiles of temperature (T), water vapormixing ratio (QVP), and wind (WSP for speed and WDR fordirection). The comparison is shown in Fig. 8 where also thebias (defined as difference between observation and model) forall variables has been computed.

Soundings launched from the center of Rome (41.90◦N,12.52◦E) during the METAWAVE campaign and from Praticadi Mare (41.65◦N, 12.43◦E) are used; the model results areinterpolated at the sites’ coordinates for the comparison. Themain differences between the two simulations in the site ofPratica di Mare are found at the start time, but no experimentaldata are available at that time (09 UTC). At 12 UTC, the twoMM5 profiles (MM5_NOVAR and MM5_VAR) are very simi-lar (not shown) and no appreciable difference is detectable forthis site.

Two soundings are available over Rome at 10 UTC and1230 UTC. The profile at 10 UTC (Fig. 8, top left panel)shows that the model overestimates the observed water vapor(gray line) near the surface, with MM5_VAR (black solid line)producing larger bias (correspondent black solid line on theleft) than MM5_NOVAR (black dashed line): an overestima-tion of approximately 1.0 g/kg (MM5_VAR) and 0.20 g/kg(MM5_NOVAR) is produced at 80 m of height. At higher lev-els, between 150 m and 1000 m, the two simulations showopposite results: MM5_NOVAR (black dashed line) producesan underestimation, whereas MM5_VAR (black solid line)an overestimation, but with a smaller bias than the controlrun. Correspondingly, a cooling of the layer is detected forMM5_VAR simulation (Fig. 8, black solid line on the topright panel) with a larger bias with respect to the obser-vations (at 80 m level the bias increases from 1.1 ◦C ofMM5_NOVAR to 1.8 ◦C of MM5_VAR). Even if MM5_VAR



Fig. 6. Water vapor mixing ratio (g/kg, solid line), temperature (◦C, dashed lines), wind vectors on a vertical cross section taken over the InSAR swath (line ABof Fig. 5) at 09 UTC of October 3rd (start time). The area inside the two vertical gray dashed-lines represents a portion of the section over the urban area of Rome(line CD of Fig. 5). The two panels refer respectively to the control run MM5-NOVAR (left) and the assimilated one MM5-VAR (right).

Fig. 7. Water vapor mixing ratio (g/kg, solid line), temperature (◦C, dashed lines), wind vectors on a vertical cross section taken over the InSAR swath (line ABof Fig. 5) at 10 UTC of October 3rd. The area inside the two vertical gray dashed-lines represents a portion of the section over the urban area of Rome (line CDof Fig. 5). The two panels refer respectively to the control run MM5-NOVAR (left) and the assimilated one MM5-VAR (right).

(Fig. 8 top right, black solid line) shows a higher tempera-ture than MM5_NOVAR (black dashed line) near the surface, itshows a larger lapse rate than MM5_NOVAR within the first50 m, resulting into an excessive cooling of the upper lay-ers. Above 1 km, both the simulations tend to overestimatethe observed water vapor (Fig. 8, top left panel) and onlynegligible differences are found between the two temperatureprofiles (Fig. 8, top right panel). On the other hand, a posi-tive impact of the InSAR assimilation is detected on the wind

fields (Fig. 8, bottom panels) with a reduction of wind speedbetween 80 and 1000 m height for the MM5_VAR simulation(black solid line). This reduces the bias (correspondent blacksolid line on the left): at 500 m, for example, a bias reduc-tion of 0.7 m/s is found. The wind direction profiles for thetwo simulations are very similar (Fig. 8, bottom right panel),but with a more marked south component of the south-westerlyflow below 250 m resulting from MM5_VAR (black solid line)than from MM5_NOVAR (black dashed line). This turns into



Fig. 8. Comparison between radio-soundings (gray lines) and model result (black dashed lines for MM5_NOVAR with its correspondent bias on the left, blacksolid lines for MM5_VAR with its correspondent bias on the left) for water vapor mixing ratio (top left), temperature (top right), wind speed (bottom left) andwind direction (bottom right) in Rome (41.90◦N, 12.52◦E) at 10 UTC of October 3rd, 2008. Biases are calculated between observed and simulated data. For eachpanel the minimum/maximum bias along the profile is indicated for the two simulations.

an enhanced advection of humid air, partially explaining themoister atmosphere at the lowest levels for this simulation.

Fig. 9 shows the vertical profiles over Rome at 1230 UTC:small differences are found between MM5_NOVAR (blackdashed lines) and MM5_VAR (black solid line). At this time,the InSAR data assimilation tends to reduce the bias for mostvariables. The mixing ratio vertical profiles show an overesti-mation for both MM5_NOVAR and MM5_VAR (Fig. 9, topleft panel black dashed and black solid lines, respectively) withrespect to radiosonde data below 1 km (bias ∼1.3–2.6 g/kg),with a very small reduction of the error below 750 m whenInSAR data are assimilated (black solid lines). Between 1.5 and3.5 km, MM5_VAR (black solid line) tends to reproduce a smo-othed profile, close to the mean of the observed profile (grayline), reducing the bias; on the other hand, the control simula-tion (black dashed line) continues to overestimate RAOB data.

Differences between MM5_NOVAR and MM5_VAR on thetemperature profiles (Fig. 9, top right panel, black dashed andblack solid lines, respectively) are small, even if reduced errors

are found when InSAR data are assimilated (black solid line).This is especially true in the layer between 1.5 and 2.5 km,where the maximum bias with respect to the observationsdecreases from 1.3 to 0.5 ◦C. This small reduction of the biasesfor both QVP and T produced by the InSAR data assimilationimproves the relative humidity profile with a reduction up to5% of the bias with respect to RAOB data below 1 km, andon average up to 10% above (between 1.5 and 3.0 km). Theimprovements produced by the InSAR data assimilation onthe water vapor content can partially be related to the correc-tion of the advection highlighted by the comparison betweenthe wind fields of the two simulations. The wind speed pro-file for MM5_VAR (Fig. 9, bottom left panel, black solid line)shows a reduction of both the overestimation with respect tothe radiosounding (gray line) below 2 km (mean bias decreasesfrom 2.4 m/s for MM5_NOVAR to 1.2 m/s for MM5_VAR)and of the underestimation between 2 and 3 km (mean biasdecreases from 1.3 m/s for MM5_NOVAR to 0.2 m/s forMM5_VAR).



Fig. 9. Comparison between radio-soundings (gray lines) and model result (black dashed lines for MM5_NOVAR with its correspondent bias on the left, blacksolid lines for MM5_VAR with its correspondent bias on the left) for water vapor mixing ratio (top left), temperature (top right), wind speed (bottom left) andwind direction (bottom right) in Rome (41.90◦N, 12.52◦E) at 1230 UTC of October 3rd, 2008. Biases are calculated between observed and simulated data. Foreach panel the minimum/maximum bias along the profile is indicated for the two simulations.

The differences of the wind direction between MM5_NOVAR and MM5_VAR (Fig. 9, bottom right panel, blackdashed and black solid lines, respectively) are small, even ifalso in this case, the InSAR data assimilation turns into a smallreduction of the bias with respect to measurements (gray line):on average from 23◦ for MM5_NOVAR (black dashed lines) to15 degrees for MM5_VAR (black solid lines).

In spite of an enhancement of the error close to the start time(10 UTC) for both the temperature and the water vapor mixingratio profiles near the surface, the results show an improve-ment of the dynamical fields that might contribute to the morecorrect evolution of the system verified with the comparisonof profiles at 1230 UTC (Fig. 9). This allows us to concludethat there is a better agreement between the assimilated sim-ulation and the observations than for the control run in termsof thermodynamical variables. Accordingly, a positive impactof the assimilation also on the precipitation forecast can behypothesized.

C. Precipitation Forecast

To assess the impact of the InSAR data assimilation on therain forecast, a comparison with the observed precipitation fieldis carried out. The rain retrieved from the Mount Midia radar(Fig. 10) is available from the Civil Protection Department ofthe Abruzzo Region [33], [34]. The radar shows rain start-ing offshore at 12 UTC and moving inland at 13 UTC (notshown). Fig. 10 shows that the 3-h accumulated rain has apattern aligned along a northeast-southwest axis (NE-SW) formost of the time. Until 15 UTC (Fig. 10, top left panel), weak tomoderate rain is detected in the southeast of Lazio: a wide areaof rainfall is shown northwest of the city of Frosinone (FR),extending up to the coast (up to 12 mm/3 h); very localizedcells are observed on the east side of Rome (RM), with rainreaching 18–20 mm/3 h (actually accumulated in 1 h between13 and 14 UTC). In the following 3 h (Fig. 10, top rightpanel), most of the Lazio region is interested by weak pre-cipitation, with intense rainfall on the east side of Rome. Two



Fig. 10. Observed 3 h rainfall estimated from Abruzzo Region Radar on Mount Midia (42.05◦N, 13.17◦E) ending at 15 UTC (top left), 18 UTC (top right), 21UTC (bottom left), 24UCT (bottom right) of October 3, 2008 over Lazio and west Abruzzo regions. Main cities of the region are indicated in pink: Rome (RM),Rieti (RT), Viterbo (VT), Latina (LT), and Frosinone (FR).

structures are detected also at this time: the first one south ofRieti (RT) with rain from 8 to 20 mm/3 h, whereas the sec-ond one reaching 18 mm/3 h with localized maxima east ofRome (also in this case, the precipitation occurred during thelast hour).

During the 3 h period ending at 21 UTC, the precipitation isspread over most of Lazio region with intense rain rates on theeast and southeast (Fig. 10, bottom left panel); hourly maps (notshown) show diffuse (3–8 mm/h) in the area around Rieti (RT)with more intense cells developing at 20 UTC (12–18 mm/h)south-east of the city. Weak precipitation (8 mm/3 h) is mea-sured in the area between Frosinone (FR) and Latina (LT),whereas spread rain falls between 20 UTC and 21 UTC onthe east side of Rome (6–10 mm/h). After 21 UTC (Fig. 10,bottom right panel), the rain moves eastward, mainly affectingthe border territories between Lazio and Abruzzo, with heavy

rain occurring between 21 UTC and 22 UTC (8–12 mm/h); rainended by midnight.

A similar rain field is found for MM5_NOVAR (Fig. 11) andMM5_VAR (Fig. 12). No influence of the InSAR data assimi-lation is found on the timing of the event: in both cases, MM5forecasts precipitation starting after 11 UTC with very weakrain rates on the east side of Rome, earlier with respect to theRadar observations. An intensification of the rain is producedafter 14 UTC. MM5 correctly reproduces the NE-SW axis ofthe rain structures, yet highlighted by the Radar measurements.

Both simulations (MM5_NOVAR and MM5_VAR) forecastthe precipitation in the Rieti district (RT) earlier than the obser-vations (before 15 UTC, Figs. 11 and 12, top left panels). In the3-h interval ending at 15 UTC, MM5 correctly reproduces twoareas of maximum precipitation [(Fig. 10, east of Rome andbetween Latina (LT) and Frosinone (FR)], in good agreement



Fig. 11. MM5 simulated 3-h rainfall without assimilation (MM5_NOVAR) ending at 15 UTC (top left), 18 UTC (top right), 21 UTC (bottom left), 24UCT (bottomright) of October 3, 2008 over Lazio and west Abruzzo regions.

with the radar, but with a displacement with respect to the obser-vations. The MM5_VAR simulation shows the first maximummore widespread than MM5_NOVAR and it produces a largeroverestimation with respect to the radar (the bias increases ofabout 8 mm/3 h).

The model reproduces the precipitation structure between LTand FR (Figs. 11 and 12, top left panel) with a westward exten-sion with respect to the radar (Fig. 10, top left panel). Bothsimulations overestimate the rainfall (Figs. 11 and 12, top leftpanels). The InSAR data assimilation partially corrects the rainintensity: the overestimation is reduced on the west side ofthe precipitation area of about 8 mm/3 h with a more realisticwest-east rain intensity gradient.

In the following 3 h (Figs. 11 and 12, top right panels), bothMM5 simulations continue to produce weak rain over Rieti(RT) district, showing a system of localized cells in partial

agreement with the radar (Fig. 10, top right panel), but none ofthem correctly reproduces the highest intensities. MM5_VAR(Fig. 12, top right panel) shows a small intensification of thecells, slightly reducing the error with respect to the observedfield. In order to explain the MM5 underestimation over Rieti(RT) of the precipitation in this time interval (15–18 UTC), onecan speculate that the early onset of the precipitation by MM5in this area excessively depletes the water vapor available forrain formation during the following hours. The InSAR assimi-lation at start time is not sufficient to fully correct this error ina few hours.

At this time (15–18 UTC), both model runs show a maximumprecipitation east of Rome (Figs. 11 and 12, top right panels).Also in this case, there is a good agreement between the modeland the radar in terms of maxima values, but MM5 producesheavy rain on a wider area than that observed; MM5_VAR,



Fig. 12. MM5 simulated 3-h rainfall with InSAR-integrated water vapor assimilation (MM5_VAR) ending at 15 UTC (top left), 18 UTC (top right), 21 UTC(bottom left), 24UCT (bottom right) of October 3, 2008 over Lazio and west Abruzzo regions.

moreover, tends to further spread the precipitation, worseningthe agreement with the observations (Fig. 12). In addition, themodel continues to produce heavy precipitation in the southernpart of the domain, largely overestimating the radar in the samearea; in this case, the InSAR data assimilation seems to havean effect on reducing the rain accumulation and the discrep-ancy with the measurements. However, the area of maximumprecipitation is too wide also in the MM5_VAR simulation.

During the successive 3 h ending at 21 UTC (Figs. 11 and12, bottom left panels), both the simulations correctly pro-duce rain over the Viterbo area (VT), slightly overestimatingthe rain retrieved by the radar. At 20 UTC, the model cor-rectly simulates the development of a few cells near Rieti (RT),with a spatial shift of the structure. The MM5_VAR simulation(Fig. 12, bottom left panel) does not correct the spatial dis-placement of the cells, but it increases the rate of the southwestcells while decreases that on the northwest side, thus partially

increasing the agreement with the radar observations. A furthersmall correction is produced by MM5_VAR reducing the rateof the cell simulated east of Rome (Figs. 11 and 12, bottom leftpanels). On the other hand, MM5 overestimates the precipita-tion on the bottom right corner of the domain by few mm/3 hup to about 9 mm/3 h for MM5_NOVAR, to about 11 mm/3 hfor MM5_VAR.

After 21 UTC (Figs. 11 and 12, bottom right panels), onlyweak rain is produced by the model, regardless of the assim-ilation process, causing a large bias with respect to the radar,partially reduced in the MM5_VAR simulation by roughly3 mm/3 h.

The results suggest that the assimilation of IWV dataretrieved from InSAR has an impact on the precipitation fore-cast, but it is not always positive. The positive impact occurswhen the rain structures develop during a time interval longerthan half an hour and spread over wide areas at a horizontal



Fig. 13. Q–Q plot of the 3h accumulated precipitation for October 3rd, 2008in the time interval between 12 and 24 UTC. Observed and forecasted quan-tile thresholds are respectively on the x and y axes. Control simulation(MM5_NOVAR) is represented in black and assimilated one (MM5_VAR)in gray.

scale comparable to or larger than that of the model. On theother hand, it fails in correcting the field, or it has even anegative impact, on very localized precipitation (model sub-grid scale). The assimilation in terms of IWV, which is a 2-Dfield, has limits in correcting the model dynamics. Significantimprovements on the rain field would be probably achievable ifwater vapor data were assimilated together with wind data [1],[31], [35]. An experiment in this sense would be very interest-ing and would probably correct at least the space bias of therain field; it is beyond the aim of this study but it represents achallenging future step.

VI. STATISTICS

To evaluate the impact of the assimilation of InSAR data, afew statistical methods and indices commonly used for weatherforecasting are applied in this section: the quantile-quantile(QQ) plot, the Equitable Threat Score (ETS), and the frequencybias (FBIAS) [36]. The QQ plot is a graphical method to com-pare the distribution of forecast and observation; data are sortedfrom smallest to largest and their percentile values are com-pared. The ETS roughly quantifies the percentage of correctforecasted rainy events that can be related to the model skill(i.e., the percentage of nonrandom correct forecasts), with val-ues ranging from slightly negative (forecast worse than random)to 1 (perfect forecast). The FBIAS score allows for evaluat-ing the frequency of the total forecasted events (hits and falsealarms) at a given threshold values, e.g., a value above/below 1indicating an over-/under-forecasted event.

The MM5 results are compared with observations from therain gauge network of the Italian Civil Protection Department(DPC) over Lazio and Abruzzo; the 145 gauges that wereavailable within the D04 domain are used for the comparison.

Fig. 13 shows the QQ plot of the 3-h accumulated rain forthe two simulations (MM5_NOVAR in black and MM5_VARin gray) with respect to the observations. MM5 produces anunderestimation of precipitation events for threshold above3 mm/3 h regardless of the assimilation process; the underesti-mation increases for medium-high threshold (>15 mm/3 h). Itis worth noting that the assimilation of the IWV retrieved fromInSAR reduces the underestimation, especially in the intervalbetween 12 and 20 mm/3 h.

The ETS is computed for 12-h accumulated rain between 21and 24 UTC of October 3 (every 3 h), with the goal of partiallyreducing the negative impact of the time bias of the event evo-lution. The ETS index increases from 0.16, 0.19, and 0.20 forMM5_NOVAR to 0.23, 0.22, and 0.23 for MM5_VAR, respec-tively, for the threshold values of 1, 3, and 6 mm. Moreover,MM5_VAR produces a higher score than MM5_NOVAR up tothe threshold of 9 mm; for intermediate thresholds (10–15 mm),the ETS decreases and differences between the two simulationsbecome negligible above 15 mm.

These results highlight that the InSAR assimilation has animpact on the forecast, with some improvements at weak pre-cipitation thresholds. This is confirmed by the FBIAS computa-tion for the 12-h accumulated rain: the MM5_VAR simulationshows an index higher than MM5_NOVAR for thresholds up to12 mm/12 h; mean FBIAS increases from 0.64 to 0.74 withinthat threshold limit. This means that the water vapor assimila-tion reduces the underestimation of the frequency of events thataffects the model for low to moderate rainfall. It is worth notingthat both the ETS and FBIAS computed for shorter accumula-tion intervals (3 h) give similar results but produce lower scores,as expected.

VII. SUMMARY AND CONCLUSION

This paper presents an experiment aimed at exploiting theAPS maps, provided by a multipass interferometric process-ing of SAR images, for the purpose of weather prediction. Inparticular, the IWV map retrieved from ASAR multipass inter-ferometric data and MERIS products has been assimilated intothe mesoscale numerical prediction model MM5.

The experiment is carried out in the framework of the ESAMETAWAVE project, as the final step of a comprehensive studyfor evaluating the water vapor path delay through the atmo-sphere and its mitigation in SAR interferometry applications.In this frame, the InSAR comes out as a potential candidate toprovide valuable information about high-resolution water vaporfield. The correct estimation of the water vapor into the weatherforecast IC is one of the most important factors for a goodforecast. A support from an external source (in this study asequence of MERIS water vapor products) is necessary to turnthe differential APS information into an absolute estimate ofthe tropospheric path delay. Once obtained, the high-resolutionwater vapor field, the data were thinned to the NWP modelresolution and assimilated using a three-dimensional (3-D)variational technique. The impact of this assimilation on theforecast is investigated by analyzing both direct and indirecteffects.



The detected differences on vertical sections of theatmosphere between the control run (MM5_NOVAR) and thesimulation with assimilated InSAR data (MM5_VAR) havehighlighted an impact of the InSAR assimilation on the modelvertical distribution of the water vapor, especially until fewhours right after the start time. The assimilation changes thethermodynamical structure of the atmosphere and it introducesa larger vertical variability of the water vapor field. The com-parison between the vertical profiles of water vapor mixingratio, temperature, and wind field shows the impact of the SARassimilation on the thermodynamical structure. It shows differ-ences between the control and assimilated simulations on thesite of Rome: despite an increase of the error by the assimilatedrun on the water vapor content and temperature at lower lev-els close to the start time, a remarkable correction of the windfield is produced by the assimilation at this time. This is sup-posed to contribute to a better forecast in the following hours,as shown by the comparison with a second sounding on thesame site at a later time, showing a better agreement with theobservations of the assimilated run than the control run for allvariables.

Finally, the impact on the precipitation forecast has beenevaluated. The model results are qualitatively compared withthe rain field retrieved from a ground-based meteorologicalradar. This comparison shows no appreciable impact of theInSAR data assimilation on the temporal evolution of the event.A positive impact would likely require the assimilation ofadditional dynamical data (i.e., wind field) in the assimilationprocess.

On the other hand, impacts on the rain intensities are found:these are positive for precipitating structures extended overwide areas (larger than the model horizontal resolution scale)and developing on time intervals longer than half an hour,whereas it is negative for convective structures at subgrid scale.Moreover, the simulation with the InSAR data assimilationimproves the forecasting performance of the spatial gradient ofthe rain, mainly, for systems with multiple cells. It is reason-able to suppose that this result could be improved by runningsimulations with resolution grid higher than 1 km, thus fullyexploiting the high resolution of the APS maps of few hundredmeters which, in principle, could provide a better description ofvery local phenomena.

A comparison between the forecasted precipitation with themeasurements available from the Civil Protection Departmentrain gauge network over the region of interest allows to assessa general underestimation of the precipitation regardless ofthe assimilation of IWV, but it also highlights a reductionof this underestimation if InSAR data are used. The equi-table threat score and frequency bias statistical indices haveshown the difficulty of the model in correctly reproducing themoderate rain for this event, regardless of the InSAR vaporassimilation. However, the improvements shown by the InSARassimilation for low precipitation thresholds (with a reductionof the model underestimation of the percentage of the totalforecasted events and the increase in the number of the pre-cipitating events correctly predicted) are encouraging for futuredevelopments.

This is the first study, to our knowledge, where differentialatmospheric delay data derived by multipass SAR interfero-metric techniques are applied for weather prediction purposes.Although these results are preliminary, given that they arededuced from only one case study, they provide the potentialof using the InSAR products in meteorological studies. Thestudy results demonstrate that the IWV properly retrieved byInSAR can be useful for mesoscale assimilation and NWP.They allow assessing some impacts of the assimilation onthe forecast, but they are not sufficient at the moment tosupport the hypothesis that such an impact is unequivocallypositive for the precipitation forecast. The results should begeneralized by adding more case studies. Other assimilationtechniques could be tested to investigate the impact of the highresolved vapor data as provided by InSAR retrieval, as well ashigh-resolution simulations. Moreover, additional advantagesderived from building optimal IC through InSAR data assimila-tion can be foreseen also by assimilating wind data from othersources.

ACKNOWLEDGMENT

The authors would like to thank all the METAWAVE partic-ipants for their valuable contribution and discussion; AbruzzoRegion in Italy for providing radar data, E. Picciotti and F. S.Marzano for their helpful suggestions; Italian Civil ProtectionDepartment for providing rain gauges data; and Universityof Wyoming Department of Atmospheric Science for makingavailable radio-soundings data on the Pratica di Mare site.

REFERENCES

[1] Y.-H. Kuo, Y.-R. Guo, and E. Westwater, “Assimilation of precipitablewater measurement into a mesoscale numerical model,” Mon. WeatherRev., vol. 121, no. 4, pp. 1215–1238, Apr. 1993.

[2] Y.-H. Kuo, X. Zou, and Y.-R. Guo, “Variational assimilation of pre-cipitable water using a nonhydrostatic mesoscale adjoint model. Part I:Moisture retrieval and sensitivity experiments,” Mon. Weather Rev.,vol. 124, no. 1, pp. 122–147, Jan. 1996.

[3] E. Pichelli et al., “Water vapour distribution at urban scale using high-resolution numerical weather model and spaceborne SAR interferometricdata,” Nat. Hazards Earth Syst. Sci., vol. 10, no. 1, pp. 121–132, Jan.2010.

[4] R. F. Hanssen, T. M. Weckwerth, H. A. Zebker, and R. Klees, “High-resolution water vapor mapping from interferometric radar measure-ments,” Science, vol. 283, no. 5406, pp. 1297–1299, Feb. 1999.

[5] G. Panegrossi, R. Ferretti, L. Pulvirenti, and N. Pierdicca, “Impact ofASAR soil moisture data on the MM5 precipitation forecast for theTanaro flood event of April 2009,” Nat. Hazards Earth Syst. Sci., vol. 11,no. 12, pp. 3135–3149, Dec. 2011.

[6] F. Rocca, “Modeling interferogram stacks,” IEEE Trans. Geosci. RemoteSens., vol. 45, no. 10, pp. 3289–3299, Oct. 2007.

[7] P. Mateus, G. Nico, and J. Catalão, “Can spaceborne SAR interferometrybe used to study the temporal evolution of PWV,” Atmos. Res., vol. 119,no. SI, pp. 70–80, Jan. 2013.

[8] N. Pierdicca et al., “Atmospheric water vapor effects on spaceborne inter-ferometric SAR imaging: Comparison with groundbased measurementsand meteorological model simulations at different scales,” in Proc. IEEEInt. Geosci. Remote Sens. Symp. (IGARSS), 2009, vol. 5, pp. 320–323.

[9] S. Sahoo et al., “3-D humidity retrieval using a network of compactmicrowave radiometers to correct for variations in wet tropospheric pathdelay in spaceborne interferometric SAR imagery,” IEEE Trans. Geosci.Remote Sens., vol. 49, no. 9, pp. 3281–3290, Sep. 2011.



[10] S. Bonafoni, V. Mattioli, P. Basili, P. Ciotti, and N. Pierdicca, “Satellite-based retrieval of precipitable water vapor over land by using a neural-network approach,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 9,pp. 3236–3248, Sep. 2011.

[11] D. Cimini et al., “On the accuracy of integrated water vapor observa-tions and the potential for mitigating electromagnetic path delay error inInSAR,” Atmos. Meas. Tech., vol. 5, no. 5, pp. 1015–1030, May 2012.

[12] M. Montopoli, N. Pierdicca, and F. S. Marzano, “Spectral downscaling ofintegrated water vapor fields from satellite infrared observations,” IEEETrans. Geosci. Remote Sens., vol. 50, no. 2, pp. 415–428, Feb. 2012.

[13] P. Basili, S. Bonafoni, P. Ciotti, and N. Pierdicca, “Modeling and sensingthe vertical structure of the atmospheric path delay by microwave radiom-etry to correct SAR interferograms,” IEEE Trans. Geosci. Remote Sens.,vol. 52, no. 2, pp. 1324–1335, Feb. 2012.

[14] A. Monti Guarnieri, S. Tebaldini, F. Rocca, and A. Broquetas, “GEMINI:Geosynchronous SAR for earth monitoring by interferometry and imag-ing,” in Proc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), 2012,pp. 210–213.

[15] A. Ferretti, C. Prati, and F. Rocca, “Permanent scatters in SAR inter-ferometry,” IEEE Trans. Geosci. Remote Sens., vol. 39, no. 1, pp. 8–20,Jan. 2001.

[16] R. F. Hanssen, “Stochastic model for radar interferometry,” in RadarInterferometry: Data Interpretation and Error Analysis. Norwell, MA,USA: Kluwer, 2001, pp. 98–110.

[17] B. Puysségur, R. Michel, and J.-P. Avouac, “Tropospheric phase delayin interferometric synthetic aperture radar estimated from meteorolog-ical model and multispectral imagery,” J. Geophys. Res., vol. 112,no. B05419, pp. 1–12, May 2007.

[18] J. Dudhia, “A nonhydrostatic version of the Penn State-NCAR MesoscaleModel: Validation tests and simulation of an Atlantic cyclone andcold front,” Mon. Weather Rev., vol. 129, no. 5, pp. 1493–1513,May 1993.

[19] R. Ferretti, G. Mastrantonio, S. Argentini, L. Santoleri, and A. Viola, “Amodel-aided investigation of winter thermally driven circulation in theItalian Tyrrhenian coast for a case study,” J. Geophys. Res., vol. 108,no. D24, pp. 4777–4792, Dec. 2003.

[20] S.-Y. Hong and H.-L. Pan, “Nonlocal boundary layer vertical diffusionin medium-range forecast model,” Mon. Weather Rev., vol. 124, no. 10,pp. 2322–2339, Oct. 1996.

[21] G. L. Stephens, “The parametrization of radiation for numerical Weatherprediction and climate models,” Mon. Weather Rev., vol. 112, no. 4,pp. 826–867, Apr. 1984.

[22] J. S. Kain and J. M. Fritsch, “Convective parameterization for mesoscalemodels: The Kain-Fritsch scheme,” Meteorol. Monogr., vol. 24, no. 46,pp. 165–170, Dec. 1993.

[23] J. S. Kain, “The Kain-Fritsch convective parameterization: An update,”J. Appl. Meteorol., vol. 43, no. 1, pp. 170–181, Jan. 2004.

[24] J. Reisner, R. Rasmussen, and R. Bruintjes, “Explicit forecasting of super-cooled liquid water in winter storms using the MM5 mesoscale model,”Quart. J. Roy. Meteorol. Soc., vol. 124, no. 548, pp. 1071–1107, Apr.1998.

[25] A. C. Lorenc et al., “The met office global three-dimensional varia-tional data assimilation scheme,” Quart. J. Roy. Meteorol. Soc., vol. 126,no. 570, pp. 2991–3012, Oct. 2000.

[26] P. Courtier et al., “The ECMWF implementation of three-dimensionalvariational assimilation (3DVAR). I: Formulation,” Quart. J. Roy.Meteorol. Soc., vol. 124, no. 550, pp. 1783–1807, Jul. 1998.

[27] D. M. Barker, W. Huang, Y.-R. Guo, A. J. Bourgeois, and Q. N. Xiao,“A three-dimensional variational data assimilation system for MM5:Implementation and initial results,” Mon. Weather Rev., vol. 132, no. 4,pp. 897–914, Apr. 2004.

[28] L. Cucurull, F. Vandenberghe, D. Barker, E. Vilaclara, and A. Rius,“Three-dimensional variational data assimilation of ground-based GPSZTD and meteorological observations during the 14 December 2001storm event over the Western Mediterranean Sea,” Mon. Weather Rev.,vol. 132, no. 3, pp. 749–763, Mar. 2004.

[29] C. Faccani et al., “Impact of a high density GPS network on theoperational forecast,” Adv. Geosci., vol. 2, pp. 73–79, Mar. 2005.

[30] D. F. Parrish and J. C. Derber, “The National meteorological center’sspectral statistical interpolation analysis system,” Mon. Weather Rev.,vol. 120, no. 8, pp. 1747–1763, Aug. 1992.

[31] M. K. Ramamurthy and F. H. Carr, “Four dimensional data assimilation inthe monsoon region. Part I: Experiments with wind data,” Mon. WeatherRev., vol. 115, no. 8, pp. 1678–1706, Aug. 1987.

[32] C. Cardinali, S. Pezzulli, and E. Andersson, “Influence matrix diagnosticof a data assimilation system,” Quart. J. Roy. Meteorol. Soc., vol. 130,no. 603, pp. 2827–2849, Oct. 2004.

[33] F. S. Marzano, E. Picciotti, and G. Vulpiani, “Rain field and reflectiv-ity vertical profile reconstruction from C-band radar volumetric data,”IEEE Trans. Geosci. Remote Sens., vol. 42, no. 5, pp. 1033–1046,May 2004.

[34] E. Picciotti, M. Montopoli, S. Di Fabio, and F. S. Marzano, “Statisticalcalibration of surface rain fields derived from C-band Mt. Midia opera-tional radar in central Italy,” in Proc. ERAD, 2010.

[35] M. K. Ramamurthy and F. H. Carr, “Four dimensional data assimilationin the monsoon region. Part II: Role of temperature and moisture data,”Mon. Weather Rev., vol. 116, no. 10, pp. 1896–1913, Oct. 1988.

[36] D. S. Wilks, Statistical Methods in the Atmospheric Science. New York,NY, USA: Academic, 1995, 467 pp.

Emanuela Pichelli received the Laurea (cum laude)and the Ph.D. degrees in physics from the Universityof L’Aquila, L’Aquila, Italy, in 2007 and 2011,respectively.

In 2011, she has a Postdoctoral Position in thegroup of weather modeling, Center of Excellence forRemote Sensing and Modeling of Severe Weather(CETEMPS), University of L’Aquila. Since 2013, shehas been a Visiting Scientist with the Mesoscale andMicroscale Meteorology Division, National Centerfor Atmospheric Research (NCAR), Boulder, CO,

USA. She participated as Investigators to National and International Projectsfunded by the Italian Civil protection Department and the European SpaceAgency. In 2012, she has been a Coorganizer of the experimental campaign overItaly of the HyMeX project. Her research interests include mesoscale meteo-rology in complex terrain areas, parameterization of turbulence in numericalweather prediction models (MM5, WRF-ARW), observations integration andassimilation into weather prediction models, and mesoscale models validationthrough observations from different source (remote, ground-based, and in situmeasurements).

Dr. Pichelli was the recipient of a Young Scientist Award EGU at the 10thPlinius Conference in 2008.

Rossella Ferretti received the Laurea (summa cum laude) in physics from theUniversity of Rome “La Sapienza,” the Ph.D. degree in geophysics from theSchool of Geophysical Sciences, Georgia Institute of Technology, Atlanta, GA,USA, and the Ph.D. degree in physics from Ministry of Education, Rome, Italy.

Since 1989, She has been working as a Researcher with the Department ofPhysics, University of L’Aquila, L’Aquila, Italy, and in 2006, she got a posi-tion as an Associate Professor with the Department of Physics, University ofL’Aquila. She is an Expert in mesoscale modelling and in charge with theReal-Time forecast, University of L’Aquila. She was a Codirector with theISSAOS Summer School on Atmospheric Data Assimilation and a Chairmanof a session on Data Assimilation at EGU 2005. She was invited at theExpert Meeting of COST 720 to organize the European Campaign LAUNCH.She was responsible for several national and international projects: RegioneAbruzzo for high-resolution weather forecast; Presidenza del Consiglio deiMinistri Dipartimento della Protezione Civile (DPC) for testing and tuninghigh-resolution weather forecast; ESA for producing high-resolution watervapor map for InSar and using via 3DVAR assimilation water vapor from InSarfor improving weather forecast. She was a National Representative for COSTAction: Action/TC/DC:ES0702. She acts as a Referee for several interna-tional journals: Meteorology and Atmospheric Physics, Atmospheric Research,NHESS, for the Research Council of Norway for a proposal on the Orographicprecipitation, and for the Netherlands e-Science Center (NLeSC) and theNetherlands Organization for Scientific Research (NWO). She is invited to jointhe Editorial Board of The Scientific World Journal. She is a Guest Editor ofQJRMET for the special issue of the Hymex campaign. She was acting as aReferee for FIRB Giovani national proposal. She was a nominated reviewerfor the Italian Super Computing Resource Allocation promoted by CINECA.She is an Author of more than 45 papers on international journals. Since1998, She has been teaching several courses such as dynamic meteorology andclimatology.

Dr. Ferretti is a Member of the Executive Committee for Implementationand Science Coordination of the HyMex project and a coordinator responsi-ble for the Italian Meteorological Center, L’Aquila, Italy, during the HyMexcampaign.



Domenico Cimini received the Laurea (cum laude)and the Ph.D. degrees from the University ofL’Aquila, L’Aquila, Italy, in 1998 and 2002, respec-tively, both in physics.

In 2002–2004, he was with the Center ofExcellence for Remote Sensing and Modeling ofSevere Weather (CETEMPS), University of L’Aquila.In 2004–2005, he was a Visiting Fellow with theCooperative Institute for Research in EnvironmentalSciences (CIRES), University of Colorado, Boulder,CO, USA. In 2005–2006, he joined with the Institute

of Methodologies for the Environmental Analysis (IMAA) of the ItalianNational Research Council (CNR) working on ground- and satellite-basedobservations of cloud properties. Since 2006, he is an Affiliate of theCenter for Environmental Technology (CET), Department of Electrical andComputer Engineering, University of Colorado, Boulder, CO, USA, wherein 2007 he served as an Adjunct Professor. He is currently with the SatelliteRemote Sensing Division, IMAA/CNR. He participated as Investigators andCoprincipal Investigators to several international projects funded by the ItalianMinistry of University and Research, the Italian Space Agency, the EuropeanSpace Agency, and the U.S. Atmospheric Radiation Measurement program.Currently, he is sharing the coordination of an International MicrowaveRadiometer Network (MWRnet), grouping more than 20 meteorological insti-tutions. Since 2002, he has been a Teaching Assistant for undergraduate andgraduate courses in remote sensing, atmospheric sounding, technologies foraviation, and hydrogeological risks assessment.

Dr. Cimini was the recipient of the Fondazione Ugo Bordoni Award 2008 inmemory of Prof. Giovanni D’Auria.

Giulia Panegrossi received the Laurea degree(Hons.) in physics from the University of Rome“La Sapienza,” Rome, Italy, and the Ph.D. degreein atmospheric sciences from the Department ofAtmospheric and Oceanic Sciences, University ofWisconsin-Madison, Madison, WI, USA, in 2004.

Since 2011, she has been a Researcher with theInstitute of Atmospheric Sciences and Climate ofthe National Research Council of Italy (ISAC-CNR),Rome, Italy. She is an Author and Coauthor of sev-eral referred publications in International Journals

and Proceedings of International Conferences. Her research interests includeremote sensing of clouds and precipitation; analysis of heavy precipitationevents through the use of NWP models (UW-NMS, MM5, WRF-ARW) andcomparison with remote, ground-based, and in situ measurements; passivemicrowave precipitation retrieval algorithms; radiative transfer through pre-cipitating clouds; microphysics characterization of precipitating clouds usingmodels and observations; development of microphysics schemes; cloud electri-fication; and nowcasting techniques.

Daniele Perissin was born in Milan, Italy, in1977. He received the Master’s degree (laurea)in telecommunications engineering and the Ph.D.degree in information technology (cum laude) fromPolitecnico di Milano, Milan, Italy, in 2002 and 2006,respectively.

He joined with the Signal Processing ResearchGroup, Politecnico di Milano, in 2002, and sincethen, he has been working on the PermanentScatterers technique in the framework of RadarRemote Sensing. In 2009, he moved to the Institute

of Space and Earth Information Science, Chinese University of Hong Kong asa Research Assistant Professor. Since October 2013, he holds a position as anAssistant Professor with the School of Civil Engineering, Purdue University(USA), West Lafayette, IN, USA. He is an Author of a patent on the use ofurban dihedral reflectors for combining multisensor Interferometric SyntheticAperture Radar (InSAR) data and he has published about 100 research worksin journals and conference proceedings. He is the Developer of the softwareSarproz for processing multitemporal InSAR data.

Dr. Perissin received the JSTARS best paper award in 2012.

Nazzareno Pierdicca (M’04–SM’13) received theLaurea (Doctor’s) degree in electronic engineering(cum laude) from the University “La Sapienza” ofRome, Rome, Italy, in 1981.

From 1978 to 1982, he worked with the ItalianAgency for Alternative Energy (ENEA). From 1982to 1990, he was working with Telespazio, Rome,Italy, in the Remote Sensing Division. In November1990, he joined with the Department of InformationEngineering, Electronics and Telecommunications,Sapienza University of Rome. He is currently a Full

Professor and teaches remote sensing, antenna, and electromagnetic fields withthe Faculty of Engineering, Sapienza University of Rome. His research inter-ests include electromagnetic scattering and emission models for sea and baresoil surfaces and their inversion, microwave radiometry of the atmosphere, andradar land applications.

Prof. Pierdicca is a past Chairman of the GRSS Central Italy Chapter.

Fabio Rocca received the Doctor Honoris Causa in geophysics from the InstitutNational Polytechnique de Lorraine, Nancy, France.

He is an Emeritus Professor of telecommunications with the Dipartimento diIngegneria Elettronica, Informazione e Bioingegneria, Politecnico di Milano,Milan, Italy. His research interests include image coding, seismic signalprocessing for oil prospections, and synthetic aperture radar.

Dr. Rocca was the President of the European Association of ExplorationGeophysicists and an Honorary Member of the Society of ExplorationGeophysicists and the European Association of Geoscientists and Engineers(EAGE). He was the recipient of the Erasmus Award from EAGE, the 2012ENI Prize for New Frontiers for Hydrocarbons, and the Chinese GovernmentInternational Scientific Technological Cooperation Award for 2014.

Bjorn Rommen received the M.S. degree in electrical engineering from theDelft University of Technology, Delft, The Netherlands, in 1999.

From 1999 onwards, he has been working in the field of Earth observa-tion with the Research and Technology Centre of the European Space Agency(ESTEC), Noordwijk, The Netherlands. Recently, he has been involved in theSentinel-1A pre-launch and post-launch activities including the commission-ing phase covering overall SAR in-orbit performance. His research interestsinclude electromagnetics theory, computational electromagnetics, and radarremote sensing.

Documents

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH ......Lafayette, IN 47907 USA (e-mail: [email protected]). N. Pierdicca is with the Department of Information Engineering, Electronics